Relational Dose-Response Modeling for Cancer Drug Studies

Wesley Tansey, Christopher Tosh, David M. Blei

Abstract: Exploratory cancer drug studies test multiple tumor cell lines against multiple candidate drugs. The goal in each paired (cell line, drug) experiment is to map out the dose-response curve of the cell line as the dose level of the drug increases. The level of natural variation and technical noise in these experiments is high, even when multiple replicates are run. Further, running all possible combinations of cell lines and drugs may be prohibitively expensive, leading to missing data. Thus, estimating the dose-response curve is a denoising and imputation task. We cast this task as a functional matrix factorization problem: finding low-dimensional structure in a matrix where every entry is a noisy function evaluated at a set of discrete points. We propose Bayesian Tensor Filtering (BTF), a hierarchical Bayesian model of matrices of functions. BTF captures the smoothness in each individual function while also being locally adaptive to sharp discontinuities. The BTF model can incorporate many types of likelihoods, making it flexible enough to handle a wide variety of data. We derive efficient Gibbs samplers for three classes of likelihoods: (i) Gaussian, for which updates are fully conjugate; (ii) binomial and related likelihoods, for which updates are conditionally conjugate through Polya-Gamma augmentation; and (iii) non-conjugate likelihoods, for which we develop an analytic truncated elliptical slice sampling routine. We compare BTF against a state-of-the-art method for dynamic Poisson matrix factorization, showing BTF better reconstructs held out data in synthetic experiments. Finally, we build a dose-response model around BTF and apply it to real data from two multi-sample, multi-drug cancer studies. We show that the BTF-based dose-response model outperforms the current standard approach in biology.

Introduction

In personalized oncology, large-scale dose-response studies are essential for identifying effective treatments. These studies involve testing numerous drug candidates against various tumor cell lines to map survival curves as a function of drug concentration. However, these experiments are often hindered by significant noise, technical artifacts (e.g., pipetting errors), and the practical impossibility of testing all possible drug-cell combinations. Traditional approaches often treat each experiment in isolation or rely on rigid parametric models that fail to capture the complex, non-stationary nature of biological responses.

Methodology: Bayesian Tensor Filtering

To address these challenges, Tansey et al. propose Bayesian Tensor Filtering (BTF), a hierarchical Bayesian framework that treats the experimental landscape as a "matrix of functions." Instead of a standard matrix of scalar values, BTF operates on a tensor where each entry is a dose-response curve.

The model leverages Functional Matrix Factorization to decompose the Cell Line × Drug × Dose tensor into low-dimensional latent factors, assuming that a few underlying biological "signatures" drive the observed responses. To balance smoothness with the ability to capture sharp transitions in drug efficacy, the authors employ Horseshoe+ priors for global-local shrinkage.

Furthermore, the authors introduce Generalized Analytic Slice Sampling (GASS), a novel MCMC routine. GASS efficiently handles non-conjugate likelihoods and linear constraints, such as ensuring survival percentages remain between 0 and 1 and maintaining monotonicity—the biological principle that increasing drug concentration should not increase cell survival.

Experimental Evaluation

The BTF framework was evaluated across three distinct settings:

  1. Synthetic Data: Benchmarking against state-of-the-art dynamic matrix factorization methods using Poisson-distributed data.
  2. Internal Pilot Study: A high-fidelity dataset from Columbia University featuring 35 drugs and 28 tumor organoids.
  3. The "Landscape" Study: A large-scale public dataset (Lee et al., 2018) encompassing 67 drugs and 284 tumor organoids.

Results and Discussion

BTF demonstrated superior performance in reconstructing held-out data compared to existing methods like Poisson matrix factorization (PGDS) and standard Logistic Factor Models (LFM). A key strength of the model is its ability to "borrow strength" across the relational structure of the data; by recognizing similarities between drugs or tumors, it can accurately impute missing or highly noisy response curves.

Dose-response curves for various tumor-drug combinations. The orange line represents the BTF model fit, with shaded regions indicating uncertainty bands.

As shown in the figure above, BTF provides robust curve estimation and meaningful uncertainty quantification. Unlike standard models that may produce overconfident but inaccurate predictions, BTF’s credible intervals effectively capture the underlying experimental variation. Additionally, the model's flexibility allows it to adapt to non-stationary jumps and plateaus that rigid parametric models often miss.

Conclusion

Bayesian Tensor Filtering represents a significant advancement in dose-response modeling by integrating relational structure with flexible functional estimation. While the current Gibbs sampling implementation is computationally intensive, future work involving Variational Inference could enable real-time application in clinical settings. By incorporating additional biological side information, such as genomic markers or chemical structures, the predictive power of BTF could be further enhanced, paving the way for more precise personalized cancer therapies.

Summary updated

Abstract

Exploratory cancer drug studies test multiple tumor cell lines against multiple candidate drugs. The goal in each paired (cell line, drug) experiment is to map out the dose-response curve of the cell line as the dose level of the drug increases. The level of natural variation and technical noise in these experiments is high, even when multiple replicates are run. Further, running all possible combinations of cell lines and drugs may be prohibitively expensive, leading to missing data. Thus, estimating the dose-response curve is a denoising and imputation task. We cast this task as a functional matrix factorization problem: finding low-dimensional structure in a matrix where every entry is a noisy function evaluated at a set of discrete points. We propose Bayesian Tensor Filtering (BTF), a hierarchical Bayesian model of matrices of functions. BTF captures the smoothness in each individual function while also being locally adaptive to sharp discontinuities. The BTF model can incorporate many types of likelihoods, making it flexible enough to handle a wide variety of data. We derive efficient Gibbs samplers for three classes of likelihoods: (i) Gaussian, for which updates are fully conjugate; (ii) binomial and related likelihoods, for which updates are conditionally conjugate through Polya-Gamma augmentation; and (iii) non-conjugate likelihoods, for which we develop an analytic truncated elliptical slice sampling routine. We compare BTF against a state-of-the-art method for dynamic Poisson matrix factorization, showing BTF better reconstructs held out data in synthetic experiments. Finally, we build a dose-response model around BTF and apply it to real data from two multi-sample, multi-drug cancer studies. We show that the BTF-based dose-response model outperforms the current standard approach in biology. Code is available at this https URL.
BibTeX
@article{tansey2019relational,
  author = {Wesley Tansey and Christopher Tosh and David M. Blei},
  title = {{Relational Dose-Response Modeling for Cancer Drug Studies}},
  journal = {The Annals of Applied Statistics},
  year = {2019},
}
Figure